coach-nemoclaw-hermes

por nvidia

Entrena a un sub-agente NemoClaw Hermes en tareas de Blender, OVRTX, OVPhysX, USD, renderizado, simulación y control en vivo de Blender. Úsalo cuando Codex deba delegar…

npx skills add https://github.com/nvidia/nemoclaw-community --skill coach-nemoclaw-hermes

Coach NemoClaw Hermes

Codex is the coach and auditor. Hermes owns Blender, OVRTX, OVPhysX, and other runtime-facing execution. Codex owns task decomposition, host/sandbox transfers, milestone receipts, artifact verification, and final reporting.

Preflight

Before delegation:

  1. Verify the NemoClaw sandbox and local inference endpoint are healthy.
  2. Verify Blender MCP and only the runtimes relevant to the task.
  3. Check for an existing Hermes request using the shared Blender or OV services. Serialize shared-runtime work; parallelize only isolated sandbox-only work with distinct inputs and outputs.
  4. Establish source ownership, a task-owned output root, and whether direct Codex-to-Blender fallback is authorized.
export PATH="$HOME/.local/bin:$PATH"
export NEMOCLAW_SANDBOX_NAME="${NEMOCLAW_SANDBOX_NAME:-ov-blender-hermes}"
nemohermes "$NEMOCLAW_SANDBOX_NAME" status
pgrep -af 'nemohermes .* exec.*(hermes|blenderraw|blenderhandoff).*chat|openshell sandbox exec.*(hermes|blenderraw|blenderhandoff).*chat' || true

Do not launch a duplicate request merely because an active request is quiet.

Execution boundaries

Keep these environments distinct:

  • Codex host: sees host repositories, source files, and host outputs.
  • Hermes sandbox: terminal and file tools see /sandbox, not host paths.
  • Blender MCP: runs inside visible host Blender and can see host paths.

Choose the topology before prompting:

  • Raw MCP: run plain hermes; the sticky blenderraw profile operates visible Blender with exploratory Blender, OVRTX, and OVPhysX tools. The explicit /sandbox/.local/bin/blenderraw alias remains available. Host paths are valid only in Blender MCP operations.
  • Typed MCP: run /sandbox/.local/bin/blenderhandoff; Hermes gets only the bounded inventory, USD, and receipt operations.
  • Sandbox-only: Codex uploads inputs before delegation and downloads outputs afterward.
  • Hybrid: Hermes inspects or exports through Blender MCP; Codex uploads the resulting host files; Hermes processes them in /sandbox; Codex downloads and validates the result.

Use the installed transfer wrapper for every crossing:

nemohermes sandbox upload "$NEMOCLAW_SANDBOX_NAME" HOST_PATH SANDBOX_PATH
nemohermes sandbox download "$NEMOCLAW_SANDBOX_NAME" SANDBOX_PATH HOST_PATH

Verify transferred file counts and SHA-256 hashes. A path in a Hermes response does not prove that a file crossed the boundary.

Coaching loop

Simple inspections or renders may be one request. For complex work, use one measurable milestone at a time, such as:

  • read-only scene inventory;
  • task-owned export with source and output hashes;
  • native render completion;
  • native OVPhysX readback;
  • replay or review capture;
  • package reopen and validation.

Each prompt must include:

  • the milestone and completion marker;
  • the relevant installed Hermes skill names to use;
  • exact host-only and sandbox-visible paths;
  • allowed execution surfaces;
  • required artifacts, evidence, and blocker behavior.

Give Hermes enough time and tool turns to reason and execute:

nemohermes "$NEMOCLAW_SANDBOX_NAME" exec --timeout 1200 -- \
  hermes chat -Q --max-turns 30 -q "$HERMES_PROMPT"

A new hermes chat -q request is a fresh conversation. Carry the prior receipt and exact transferred paths into the next milestone. Poll an existing command session until it exits; logs are progress evidence, not a completion receipt.

Intervention policy

Silence alone is not failure. Intervene only when evidence shows:

  • an explicit tool or runtime error;
  • a wrong filesystem or execution boundary;
  • a missed acceptance criterion;
  • a receipt contradicted by artifacts;
  • a dropped/truncated tool call or other known agent-stream failure signature.

Diagnose the narrow layer first: wrapper, transfer, sandbox, Blender MCP, Blender operation, OVRTX/OVPhysX runtime, or artifact validation. Issue at most one targeted corrected retry before reporting the blocker. Do not repeat a broad prompt or casually overwrite artifacts still open in Blender or PXR.

Do not bypass Hermes with direct Codex control of Blender unless the user explicitly authorizes fallback execution. When authorized, report the fallback and preserve the same source-safety and evidence requirements.

Validation

Preserve source scenes and source USD layers by default. Keep generated work under the task root and hash source files before and after mutation-capable milestones.

Independently validate downloaded host artifacts. Distinguish evidence:

  • native OVRTX or OVPhysX receipts/readback are runtime authority;
  • Blender viewport images and replay renders are visual evidence;
  • Hermes text without matching artifacts is an unverified claim.

Report separately what Codex coached and verified, what Hermes executed, all artifact paths, versions, elapsed time, validation results, fallbacks, and remaining blockers or limitations.

Más skills de nvidia

compileiq-debug
nvidia
Úsalo cuando algo esté mal: Search() se cuelga, todas las evaluaciones devuelven INVALID_SCORE, las puntuaciones no mejoran, cada configuración devuelve el mismo número, errores de ptxas…
create-github-pr
nvidia
Crear solicitudes de extracción de GitHub usando la CLI gh. Usar cuando el usuario quiera crear un nuevo PR, enviar código para revisión o abrir una solicitud de extracción. Palabras clave de activación -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
Escanea otros issues abiertos para encontrar aquellos que un PR dado también podría corregir o romper accidentalmente. Genera oportunidades de corrección adyacente y riesgos de contradicción con archivo:línea…
fhir-basics
nvidia
Enseña a los agentes cómo funcionan las APIs de FHIR R4, qué recursos están disponibles, cómo consultarlos con parámetros de búsqueda y cómo analizar correctamente todos los formatos de respuesta…
compileiq-validate-result
nvidia
Usar DESPUÉS de que una Búsqueda haya finalizado y ANTES de reclamar cualquier aceleración o enviar un ACF. Carga el CSV de dump_results, extrae los mejores K candidatos (de un solo objetivo)…
changelog-audit
nvidia
Auditar el CHANGELOG.md de Warp antes de un lanzamiento: recuperar entradas perdidas, ordenar por impacto en el usuario, refinar el lenguaje de las entradas, ajustar saltos de línea y (en modo rama de lanzamiento) incrementar comparación…
maintain-dynamic-plugins
nvidia
Mantener los cargadores de plugins dinámicos de NeMo Relay, manifiestos, SDKs nativos de Rust, protocolo de trabajador gRPC, SDK de trabajador Python, documentación, pruebas y cobertura del flujo de trabajo de lanzamiento
dgx-diagnose
nvidia
Diagnostica problemas comunes de la DGX Station GB300: fallos de CUDA, direccionamiento incorrecto de GPU, errores de contenedores vLLM/SGLang, problemas de estado MIG, errores de NVLink/Fabric Manager,…